Reflexion

Stores the full trajectory of a trial, including the initial legal reasoning trace, the generated output, and the external evaluator's linguistic feedback. This persistent memory allows the agent to reference past failures in-context during subsequent trials. Unlike standard short-term memory, this buffer retains the specific error context—such as a misinterpreted statutory clause—enabling the model to learn from its own history within the active context window.

The core mechanism where an external evaluator—which can be a separate LLM, a heuristic function, or a symbolic parser—generates natural language critiques of the agent's output. For legal tasks, this feedback is highly specific:

• Identifies a hallucinated case citation
• Points to a violated canon of statutory construction
• Flags a logical gap in a syllogism This textual critique is appended to the prompt for the next trial, guiding the agent toward a corrected response.

When an external evaluator is unavailable, the agent can generate its own feedback using a chain-of-verification approach. The model is prompted to:

1. Decompose its own legal conclusion into discrete factual claims
2. Draft independent verification questions for each claim
3. Answer those questions against the provided evidence This self-generated critique serves as the linguistic signal for the next iteration, enabling autonomous improvement in environments without ground-truth labels.

A recency-focused memory management technique that retains the most recent 'N' trials within the model's active context. Older, less relevant failures are dropped to prevent the prompt from exceeding the model's token limit. This sliding window ensures the agent focuses on its most recent mistakes—critical for long chains of legal reasoning where early errors may be rendered irrelevant by later corrections in the iterative process.

Reflexion tightly couples a specific action (the generated legal text) with its outcome (the evaluator's feedback). This pairing is stored as a structured tuple in the episodic buffer. For example:

• Action: 'The contract is void under the Statute of Frauds.'
• Outcome: 'Error: The cited statute applies only to contracts for the sale of goods over $500. The contract in question is for services.' This explicit coupling allows the model to learn a direct mapping between a flawed reasoning step and its correction.

To prevent the agent from getting stuck in a local minimum of flawed logic, Reflexion can inject controlled randomness into the reasoning process. This is achieved by prompting the model to generate multiple diverse reasoning paths (similar to Tree-of-Thoughts) before the evaluator selects the most promising one. This balances exploitation of known good heuristics with exploration of novel legal interpretations, crucial for navigating ambiguous case law where a single precedent may be misleading.